1. Introduction to single-cell RNA-seq
2. Quality control and normalization
3. Survey of downstream analysis methodology

• standard QC on raw reads (adapter/quality trim)
• mapping to reference genome - bulk tools appropriate (e.g. Salmon, STAR, RSEM, etc)
• count reads per transcript/gene - bulk tools appropriate (e.g. Subread)
• collapse UMIs & demultiplex (if applicable) -> scPipe

Amezquita et al. 2019 (https://doi.org/10.1101/590562)

## class: SingleCellExperiment 
## dim: 23918 192 
## metadata(0):
## assays(1): counts
## rownames(23918): ENSMUSG00000103377 ENSMUSG00000103147 ...
##   ERCC-00171 CBFB-MYH11-mcherry
## rowData names(13): GeneLength ENSEMBL ... total_counts
##   log10_total_counts
## colnames(192): SLX-9555.N701_S502.C89V9ANXX.s_1.r_1
##   SLX-9555.N701_S503.C89V9ANXX.s_1.r_1 ...
##   SLX-11312.N712_S508.H5H5YBBXX.s_8.r_1
##   SLX-11312.N712_S517.H5H5YBBXX.s_8.r_1
## colData names(48): Plate Oncogene ...
##   pct_counts_in_top_500_features_ERCC PlateOnco
## reducedDimNames(0):
## spikeNames(1): ERCC

This SingleCellExperiment object has 23918 genes and 192 cells.

• library size (= total number of reads/UMIs)
• number of expressed features / detection rate / dropout
• expression levels by gene
• proportion of reads mapped to spike-ins (if available) or MT genes
• [Droplet] remove empty droplets
• [Droplet] remove doublets

• Ideally each droplet contains one cell and one bead
• Process of adding cells (and beads) to droplets is stochastic, approximated by a Poisson process: \[ P(k \text{ cells in a droplet}) = \frac{e^{-\lambda}\lambda^k}{k!}\]
• Typically cells are added at a low rate (\(\lambda\) << 1) to reduce the chance that a droplet contains multiple cells
• Unfortunately, this means many droplets contain zero cells
• Also unfortunately, reads still map to droplets with no cells

Lun et al. 2019, (https://doi.org/10.1186/s13059-019-1662-y)

• e.g. Quantile (CellRanger), Knee plot (Drop-seq)
• This risks removing cells with low biological material

• Lun et al. 2019, (https://doi.org/10.1186/s13059-019-1662-y)
• Key idea: expression profile of empty droplets \(\sim\) ambient RNA - Estimate ambient RNA profile by pooling across all cells/barcodes
• Ambient expression \(A_g\) for gene \(g\) is sum of counts \(y_{bg}\) over all barcodes \(b\): \[ A_g = \sum_b y_{bg} \]
• Counts for barcode \(b\) conditional on total counts for barcode \(b\) ~ Dirichlet-multinomial
  • allows for overdispersion (analogous to Gamma-poisson)
• Barcodes with strong deviations (permutation test on likelihood) are determined to be cells

• Even with the low cell rate, occasionally droplets contain two or more cells
• Can also happen in plate-based protocols (though rare) due to improper cell sorting/capture
• Simple idea: remove barcodes with especially high numbers of reads - works poorly with diverse cell types

• Bach et al. 2017 (https://doi.org/10.1038/s41467-017-02001-5)
• Key idea: remove cells that look like a mixture of two clusters

• Scrublet: Wolock, Lopez & Klein 2019 (https://doi.org/10.1016/j.cels.2018.11.005)
• DoubletFinder: McGinnis, Murrow & Gartner 2019 (https://doi.org/10.1016/j.cels.2019.03.003)
• Dahlin et al. 2018 (https://doi.org/10.1182/blood-2017-12-821413)
• Key idea: remove cells that look like simulated doublets

• Now that we’ve removed problematic cells, need to put counts on a comparable scale before/during downstream analysis
• Main factors to account for:
  • library size / sequencing depth
  • batch effects (known or unknown)

• DESeq2 library size factors use the median ratio method - Anders and Huber 2010 (https://doi.org/10.1186/gb-2010-11-10-r106): \[ \hat{s}_b = \text{median}_g \frac{y_{bg}}{(\prod_{j=1}^B y_{gj})^{1/B}} \]
• assumes majority of genes are not differentially expressed

• Consider this quantity in single-cell: \[ \hat{s}_b = \text{median}_g \color{red}{\frac{y_{bg}}{(\prod_{j=1}^B y_{gj})^{1/B}}} \]
• For every gene \(g\) where \(y_{bg} = 0\) for any \(b\), the highlighted quantity will be undefined
• This means that the quantity is only computed for genes that are nonzero in every cell
• As a result, log CPM normalization commonly used for single-cell: \[ log( \frac{y_{bg}*10^6}{\sum_b y_{bg}} + 1) \]

• scran: Lun et al. 2016 (https://doi.org/10.1186/s13059-016-0947-7) proposed a pooling & deconvolution approach

• Cells are pooled together and normalized against a global pseudoreference, then deconvolved by solving a system of linear equations

• SCnorm: Bacher et al. 2017 (https://doi.org/10.1038/nmeth.4263)
• Key idea: groups of genes get separate size factors based on observed count-depth relationship
• Ignores zeroes (developed for plate-based read count data)

Bacher et al. 2017 (https://doi.org/10.1038/nmeth.4263)

• scTransform: Hafemeister & Satija 2019 (https://doi.org/10.1101/576827)
• Fits negative binomial generalized linear model for groups of genes with similar geometric means: \[ log(E(y_{bg})) = \beta_0 + \beta_1 log_{10}(t_b), \] where \(t_b = \sum_g y_{bg}\) is the library size for cell \(b\)
• Key idea: normalized counts are taken as residuals of the above model (‘regressing out’ library size)

Hafemeister & Satija 2019 (https://doi.org/10.1101/576827)

• When batches are known and cell populations are identical across batches \(z_b\) (e.g. evenly split across two plates & homogeneous), can estimate effects due to batch

• limma: Ritchie et al. 2015 (https://doi/org/10.1093/nar/gkv007)
  • fit a linear model to log-normalized counts to estimate batch effects \(\beta_2\) and effects of interest \(\beta_1\): \[ log(y_{bg}) = \beta_0 + \beta_1 x_b + \beta_2 z_{b} \]
  • remove batch effect: \[ log(y_{bg}) - \beta_2 z_{b} \]

• see also Combat: Johnson & Rabinovic 2007 (https://doi.org/10.1093/biostatistics/kxj037), implemented in sva

• When batch composition is unknown/variable, can’t estimate common batch effect

- MNNcorrect: Haghverdi et al. 2018 (https://doi.org/10.1038/nbt.4091) aligns cells based on their mutual nearest neigbors

• See also Seurat: Butler et al. 2018 (https://doi.org/10.1038/nbt.4096)

Haghverdi et al. 2018 (https://doi.org/10.1038/nbt.4091)

• Review: a VST removes the dependence of the variance on the mean
• It’s useful to a apply variance stabilizing transformation before:
  • visualization
  • selecting highly variable genes
  • using downstream analysis tools that don’t explicitly model count data (e.g. MAST, scDD)
• Examples:
  • log (ideal when relative differences are important rather than absolute; requires addition of a pseudocount)
  • DESeq2::vst (estimates and removes trend of mean and variance)
• Not necessary if data was already log CPM normalized

Andrews & Hemburg 2019 (https://doi.org/10.12688/f1000research.16613.1)

• scater: visualization, quality control (Bioconductor)
• scran: normalization, doublet detection, batch effect correction (Bioconductor)
• SCnorm: normalization (Bioconductor)
• sctransform: normalization (CRAN)
• DropletUtils: removal of empty droplets (Bioconductor)
• Seurat: normalization (CRAN)